美国人工智能战略内幕：基础设施、监管与全球竞争

很高兴见到大家，我非常兴奋能探讨今天的话题——人工智能和AI在我们世界中的应用。

Great to see everyone, and I'm thrilled to be able to talk about the issue of the day, and that is artificial intelligence and AI in our world.

大卫、迈克尔，我很想听听你们对当前我们在争夺全球头号AI强国地位方面的进展的看法。

David, Michael, I'd love you to talk about what where we are right now in terms of the pursuit to be the number one lead AI country.

大卫，我们目前做得怎么样？

How are we doing, David?

我觉得我们做得非常好。

I think we're doing great.

玛丽亚，去年特朗普总统发表了一次重要的AI政策演讲。

Maria, last year, president Trump gave a major AI policy speech.

那是今年七月，他宣布美国必须赢得人工智能竞赛。

This is in July, and he declared that The United States had to win the AI race.

他首先宣布，我们已经置身于一场竞赛之中。

He he had, first of all, declared that we were in one.

我认为他的演讲让人联想到肯尼迪总统宣布我们正在参与太空竞赛并必须赢得这场竞赛的时刻。

And I think his speech was reminiscent of when president Kennedy declared that we were in a space race and had to win that race.

我认为从那以后，你看到的是美国公司更加积极地创新。

I think since then, what you've seen is that American companies have only innovated more.

你不断看到各种令人惊叹的产品陆续发布。

You're seeing all sorts of really incredible products being released all the time.

我认为美国的AI模型、芯片和数据中心一直在持续进步。

I think that American AI models, chips, data centers only just keep getting better and better.

因此，我对美国在这一AI竞赛中的地位感到非常乐观。

So I feel very good about the American position in this AI race.

当然，我们有一些非常有能力且强大的竞争对手。

Certainly we have some very competent and formidable competitors.

中国显然有很多非常聪明的人在这一领域工作。

China obviously has a lot of very smart people working in this area.

但我确实认为，如今硅谷的美国公司所展现出的成就真的令人惊叹。

But I do think that just what you see from American companies in Silicon Valley right now is really incredible.

然而，关于为建设数据中心而进行的巨额支出，仍然存在许多问题。

And yet there are still so many questions about all of the spending underway to build this out with regard to data centers.

当然，人们不断提出一个问题：我们是不是花得太多了？

And of course, the question keeps coming up, are we spending too much?

我们能获得投资回报吗？

Will we get the return on investment?

你怎么看这个问题？

How do you see that?

我认为我们会的。

I I think that we will.

我认为之所以看到如此大规模的基础设施建设，是因为最终需求是真实存在的。

I think that the reason why seeing this huge infrastructure build out is because the demand is ultimately there.

我知道很多人担心这会不会像互联网泡沫那样。

I know a lot of people worry about whether this could be like a .com situation.

记得九十年代末我们曾大规模建设光纤网络，随后发生了互联网泡沫破裂。

Remember where we had the whole fiber build out in the late nineties, then we had a .com crash.

这里的不同在于，九十年代末和两千年初，我们遇到了所谓的‘暗光纤’问题，即光纤网络建好了却未被使用。

The difference here is that in the late nineties and early two thousands, we had a problem known as dark fiber where you had this fiber build out and then it didn't get used.

现在根本不存在闲置的GPU。

There's no such thing as a dark GPU right now.

每个被部署在数据中心的GPU都在被使用。

Every GPU that's being put in a data center is getting used.

而且它们被用来生成令牌。

And it's being used to generate tokens.

这些令牌用于驱动新一代的AI聊天机器人或编程助手。

And that's to power this new generation of AI chatbots or coding assistants.

在过去几个月里，编程领域有一些新发布，如果你关注软件开发者的看法，他们会说这令人震惊。

And there's just been some releases in the last couple of months on the coding front that if you're following what software developers are saying, they're saying it's mind blowing.

这完全改变了他们的行业。

It's completely revolutionizing their industry.

因此，对令牌的需求不断增长，这也推动了我们正在目睹的数据中心建设需求。

So demand for tokens just increases and that increases the demand for this data center build out that we're seeing.

所以我认为这种趋势短期内不会停止。

So I don't think it's going to stop anytime soon.

就在去年，这一基础设施建设为GDP增长率贡献了约2%。

And just last year, this infrastructure build out added about 2% to the GDP growth rate.

我认为，这正是推动我们达到四到五个百分点增长率的关键因素。

And I think and that's what helped propel us to this, you know, four to 5% growth rate.

我认为今年你也会看到类似的情况。

And I think you're gonna see something similar this year.

这确实引领了增长，迈克尔。

Well, it is certainly leading growth, Michael.

我很高兴能与你们两位——真正引领这一趋势的人——展开这场对话。

And I'm so happy to be able to get this conversation going with both of you who are really leading this.

大卫，谢谢您。

David, thank you.

迈克尔，谢谢您。

And Michael, thank you.

迈克尔，我也想问你同样的问题。

Same questions for you, Michael.

评估一下我们目前在人工智能方面的状况。

Assess where we are right now on AI.

我觉得

I I think

提醒一下大家，对于那些没有像我们一样每天密切跟踪的人，这个计划实际上有三个支柱。

just a reminder for the group for those who haven't been tracking as closely as we do every day, the the the plan really had essentially three pillars.

第一，美国如何继续超越我们的竞争对手？

And it talked about how, one, how can The US continue to out innovate our competitors?

第二，我们如何推动建设支持这场人工智能革命所需的基础设施？

Two, how can we drive the infrastructure build that we need to support this this AI revolution?

第三，我们如何向世界分享或出口我们的先进美国技术？

And three, how do we actually share with the world or export our great American technology?

针对这三个支柱，联邦政府采取了大量措施来推动其进展。

And for each of those three pillars, there was quite a lot of actions that the federal government has taken to drive that forward.

我认为，我们可以很自豪地说，我们在所有这三个方面都取得了相当不错的进展。

And I think I think we're pretty proud to say that we've made, I think, pretty good progress on on all three.

我们稍微聚焦一下你之前提到的创新问题，我认为我们一直以来的核心见解是，要推动这种创新，就必须营造一个允许这项技术在美国开发并最终商业化的监管环境。

Just focusing a little bit on the innovation one you're talking about earlier, I think the the the the the core insight that we've always had about how you drive this innovation is you have to have a regulatory environment that allows this technology to be developed and ultimately commercialized in The United States.

与世界其他地区相比，美国在这方面做得非常好，建立了一套行之有效的框架，但我们仍可以做得更好、不断改进。

And US has done a a great job compared to the rest of the world on sort of setting that up and creating a framework that works, but we could always do better and improve it.

总统在今年七月的演讲中，大量谈到了各州监管碎片化的问题，以及如何确保不会出现50种不同的AI规则。

And the president, his his speech in July, talked a lot about this issue of a patchwork of state regulations and how can we ensure that there aren't 50 different rules around AI.

这场辩论中最重要的一点，我认为很多人有时会忽略，那就是监管碎片化实际上对初创公司和创业者伤害最大。

And and what's most what's most important about this debate, which I I think a lot of people sometimes don't sometimes miss, is the patchwork is actually most detrimental to early stage young companies and entrepreneurs.

如果你想要开发一种新的AI技术，或者基于我们顶尖的前沿模型构建产品，却要应对50个州的50套不同规则，这会制造巨大的障碍。

If you wanna develop a new AI technology, if you wanna build something on top of one of our great frontier models, having to figure out how to navigate 50 different rules across 50 different states creates a lot of friction.

最终，只有大公司才能在这种环境中取得最佳优势。

And ultimately, big guys are the ones that can succeed in that environment the best.

因此，我们花了很多时间思考，如何制定一项立法提案，真正建立起一套合理且可行的国家框架，以解决这一监管难题。

So we're spending a lot of time trying to think about how can you create a legislative proposal that can actually deliver on a sensible national framework to solve to solve this regulatory issue.

那么，迈克尔，你认为这种联邦监管框架中必须包含哪些基本要素？

So so what would you say then, Michael, are the basic frameworks that are sort of must have in in that kind of federal oversight?

因为一些州在美国曾表示反对，说不行。

Because some states did push back in The US and say, no.

不行。

No.

不行。

No.

我们希望在人工智能方面能够掌控自己的命运。

We wanna be able to control our destiny when it comes to AI.

当你审视这种联邦监管框架时，最重要的是什么？

What's most important when you look at that framework in terms of a federal oversight?

是的。

Yeah.

我认为总统在去年12月签署的行政命令中，要求我们推进这项提案，并列出了一些各州仍可自行推进的事项。

I think in the executive order that the president signed in in December directing us to kinda work through this proposal, he listed a few things that the state should continue to be able to pursue individually on their own.

儿童安全方面的立法或规定就在那张清单上。

Legislation or rules around child safety was on that list.

关于数据中心建设和扩展的许可规则，仍然是各州应该关注的问题。

The rules around permitting of data centers and build outs are continuing to be something that states should should look at.

因此，有一些被明确列出的事项，但我想戴夫和我接下来要处理的就是这类问题。

So there are a few things that were enumerated, but that's the kind of stuff that I guess Dave and I are gonna be working through.

我不知道你对这方面有什么看法。

I don't know if you have any thoughts on on that.

是的。

Yeah.

我的意思是，我们面临的基本问题是，坦率地说，各州现在正在疯狂地推进监管。

I mean, I I think the the basic problem that we have is that, I mean frankly the states are going hog wild right now with regulation.

目前有超过1200项法案正在各州立法机构中审议。

There's over 1,200 bills going through state legislatures right now.

我认为这很大程度上是一种本能反应。

I think it's very much a knee jerk reaction.

我知道人们对人工智能有很多恐惧和担忧。

I know there's a lot of fears and concerns about AI.

但似乎每个假设性的担忧，现在都有多个州的法案试图在我们真正了解其发展之前就对其进行监管。

But it seems like for every hypothetical concern, there's multiple state bills now to try and regulate that thing before we really know how it's gonna play out.

我认为更好的做法是，鉴于这项技术如此新颖，环境又如此动态，我们应花更多时间研究AI实际是如何被使用的，以及哪些风险真正出现了，然后再进行过度监管。

And I think it would be better to, I think since this technology is so new and the environment is so dynamic, I think it'd be better to spend a little bit more time studying how AI is actually being used and what risks are actually materializing before you overregulate the thing.

但无论如何，这正是我们目前在州一级所看到的情况。

But in any event, that that's the what we're seeing right now at the state level.

我认为总统一直持一致立场，认为最好制定一个统一的规则手册，即一个联邦层面的轻量级标准。

And I think that the president's been very consistent that it would be better to have a single, we'll have one rulebook, a single rulebook at the federal level, lightweight federal standard.

我认为这个问题随着时间推移只会变得更加紧迫，因为当50个州各自朝不同方向前进时，这种碎片化问题只会越来越严重。

I think this problem's only gonna get more acute over time because again, as you have 50 different states running in 50 different directions, the patchwork problem only gets more significant.

所以无论如何，今年我们将密切合作，看看能否就联邦框架达成足够共识，从而推动立法。

So in any event, this is something that we're gonna work, I I think, closely together on this year, which is to see if we can get enough consensus on a federal framework to enact a law.

只有国会才能最终取代各州的法规。

Only Congress can ultimately preempt the states.

我们理解这一点。

We understand that.

你知道，正如你所知，要在国会通过一项法案非常困难。

And you know, as you know, it's very difficult to get a bill through Congress.

参议院需要60票才能通过。

You need 60 votes in the Senate.

因此，它在一定程度上必须是两党共识的产物。

So it has to be bipartisan to to a certain degree.

但我们会努力看看能否达成这种共识。

So but we're gonna try and see if we can work to get that consensus.

是的。

Yeah.

关于国会支持联邦监管的时间表，你有任何明确的信息吗？

And do you have any clarity on the timing on that in terms of support in Congress for a federal oversight?

还是说，根据你所讨论的州，你也看到了类似的阻力？

Or do you see pushback there as well, depending on the state you're talking about?

嗯，在国会内部，对于在没有联邦标准的情况下实施 preemptive 措施，确实存在阻力。

Well, there's pushback in congress to the idea of preemption without a a federal standard.

换句话说，你不能用空无代替某物。

So in other words, you can't replace something with nothing.

这正是我们反复听到的说法。

This is sort of the the thing that we heard repeatedly.

但我认为，众议院和参议院都有相当大的兴趣，希望确立某种轻量级的联邦标准。

But I think there is quite a bit of interest in both the House and the Senate towards having, again, some sort of lightweight federal standard.

但我们仍处于这些讨论的初期阶段，我们会看看今年能推动实现哪些成果。

But we're still in the early stages of those conversations, and we're gonna see what we can try and get done this year.

与此同时，一些原本推动数据中心创新与增长的人现在开始反对了。

Meanwhile, you've got some people pushing back after wanting to see the innovation and growth of data centers.

现在他们说：别建在我家后院。

Now they're saying, not in my backyard.

对此你怎么看？

What about that?

这是个问题吗？

Is that an issue?

是的。

Yeah.

我的意思是，我们最近收到了伯尼·桑德斯的一封信，说要停止所有数据中心和数据中心的发展。

I mean, we got a letter recently from Bernie Sanders saying, stop all data centers, all data center development.

如果我们这么做，就会输掉人工智能竞赛。

And, you know, if we do that, we will lose the AI race.

我的意思是，你确实需要这样的基础设施。

I mean, you do need this infrastructure.

其他国家正在建设这种基础设施。

Other countries are building out this infrastructure.

中国正在建设中。

China's building out.

我认为他们每周都在新建一座核电厂或煤电厂，或者新的能源设施。

I think they're spinning up a new nuclear power plant or or coal plant, new energy every single week.

其中很大一部分将用于为他们的数据中心供电。

And a lot of that's going to power their data center.

所以，如果我们完全停止建设数据中心，我认为这将从根本上削弱美国在人工智能竞赛中的地位。

So it would fundamentally, I think, cripple The United States in the AI AI race if we just stop building data centers altogether.

同时，人们也担心电价的可负担性，以及消费者是否会因为数据中心而不得不支付更高的电费。

At the same time, there are concerns about affordability, about whether consumers would have to pay a higher electrical rate because of data centers.

特朗普总统一直明确表示，消费者不应因为数据中心而承担更高的电费。

President Trump's been really clear that consumers should not have to pay higher rates for electricity because of data centers.

就在上周，微软率先承诺，其数据中心不会导致居民电价上涨。

You saw just last week, Microsoft stepped up and made a pledge that it will that its data centers will not cause residential rates to increase.

我认为，其他科技公司很可能会跟进，做出类似的承诺。

I think you'll likely see other tech companies stepping up and making similar commitments.

事实上，当我与超大规模云服务商或人工智能公司交谈时，他们从未计划从电网中大量抽电。

And in fact, when I've talked to the hyperscalers or when I've talked to the AI companies, it was never their plan to draw off their grid.

他们所有人都计划在建设过程中自建发电设施。

They all are saw standing up their own power generation as part of their build out.

能源部长赖特一直在努力改革那些实际上使这些人工智能数据中心难以自建电力设施的监管规定。

And what Secretary Wright, Secretary of Energy has been doing is trying to is reform the regulations that actually make it more difficult for these AI data centers to stand up their own power behind the meter.

所以这基本上就是我们的愿景，我应该说，这实际上是特朗普总统从本届政府一开始就提出的愿景：让人工智能公司成为电力公司。

So that basically is is our vision, is let and and I should say this is president Trump's vision really since the beginning of the administration, is he said, let the AI companies become power companies.

让它们在建设这些新数据中心的同时，自行建立电力供应系统。

Let them stand up their own power generation as they built, you know, side by side with these new data centers.

这样做的结果是，我们能够获得这样的基础设施。

And the the result of that is, you know, a, we get this infrastructure.

其次，居民电价不会上涨。

B, residential rates don't go up.

是的。

Yeah.

因为，迈克尔，这场竞赛已经迅速从人工智能竞赛转变为电力竞赛。

Because, Michael, the this this race has fast become its moves from an AI race to a power race.

嗯。

Mhmm.

嗯。

Mhmm.

而且我认为，我们正在看到的是，我们需要讲好一个故事，说明最终这一基础设施建设将为美国的电力用户带来净收益。

And and I think what we're seeing is that we we need to share a good story about how ultimately this build out is gonna be net positive for American ratepayers.

我认为，有时候如果你身处一个小型社区，有人来建设数据中心，你就必须明确说明，从长远来看，这实际上会降低你的电费。

And I think sometimes if, you know, and if you're in a small community and someone shows up to build a data center, I mean you have to make it clear that ultimately this is something that's gonna actually lower your your rates long term.

而且总统上星期一发布了一项声明，正如大卫所说，他明确表示，如果你要建设数据中心，就必须自己承担成本。

And and the president put out a truth last Monday where he was, as as David said, very clear that if you're gonna build a data center, have to pay your own way for it.

微软已经采取了行动，我们希望许多其他公司也能效仿。

And Microsoft has stepped up and our our hope is that many others will do the same.

但有些公司因为目前没有现金，正在借钱来建设数据中心，对吧？

But but some companies, because they don't have the cash right now, are borrowing money, right, to to build out the data centers.

人们也担心银行最终会为其中一部分支出买单，因为开支实在太大了。

And there's also a worry that the banks will be left holding the bag for some of this because, again, the spending is is too much.

你对这个问题有什么看法？

Your thoughts on that.

嗯，我认为确实存在这样的担忧。

Well, I think there there is obviously that concern.

我的意思是，我认为这与其说是银行的问题，不如说你看到甲骨文正在做出巨大投资。

I mean, you know, I I think it's it's it's less, I would say, the banks are more you see Oracle making a huge investment.

你看到黑石集团在进行巨额投资，还有房地产公司。

You see Blackstone making huge investments, real estate companies.

最终，我认为这些是非常精明的市场参与者，资金雄厚的公司，他们这么做是因为看到了彩虹尽头的投资回报。

Ultimately, I think these are very savvy market players, very deep pocketed companies, and they're doing this because they see an ROI there at the end of the the rainbow.

我能不能再提一点关于数据中心的事？

Can I just make one other point about just the the data center?

就电力而言，我认为如果我们允许数据中心自行发电，实际上会降低电价。

So just on electricity, I actually think that if we allow the data centers to stand up their own power generation, it will actually bring down rates.

这不仅不会增加居民电价，反而会降低电价，而且会通过两种方式实现。

Not only will it not increase residential rates, it'll bring it down, and it'll do that in two ways.

一是数据中心在有富余电力时可以向电网输送或出售电力。

One is that the data centers can can give or sell power back to the meter when they have excess.

这将有助于降低电价。

So that will help bring down rates.

其次，电力生产涉及大量固定成本。

Second, there's a lot of fixed costs involved in power generation.

并非全部都是可变成本。

It's not all variable.

因此，当你能将这些固定成本分摊到更大的供电规模上时，就能降低每个人的电费单价。

So when you're able to amortize those fixed costs over a greater supply, you bring down the meter rate for everybody.

因此，这里存在巨大的规模经济效应。

And so there's huge economies of scale.

所以，电力的规模越大，就像其他大多数事物一样，价格就会下降。

So the more scale you get in electricity, like most other things, the price comes down.

因此，我们正在推进的这种建设实际上是一件好事，因为它最终会降低消费者的电价。

That's what so it's actually a good thing that the that we have this build out going on because it will ultimately reduce prices for consumers.

但我们必须确保这些新的数据中心不仅仅是接入电网并消耗电力。

But we do have to make sure that these new data centers aren't just plugging into the grid and using.

它们必须也要反哺电网。

They have to be contributing back.

我认为拜登政府的一项重要政策变革是，明确禁止了这种‘电表后’的发电方式。

And I think what a a great policy changes made under this administration, the the Biden administration had, as a matter of policy, made it such that you couldn't do this behind the meter energy generation.

如果你想要自己发电，那是不允许的。

If you wanted to bring your own power, you couldn't.

你必须接入更大的电网。

You had to be part of the larger grid.

我认为，这一规定已经由怀特部长和联邦能源监管委员会修改，允许这种做法得以实现。

So I think that rule has has changed by by secretary Wright and by FERC to kinda allow this to happen.

最终，我同意戴维的观点。

And ultimately, I agree with David.

我认为，一旦电力生产达到更大的规模，你就会以一种有利于用户的反向方式向电网回馈电力。

I think once you have sort of greater scale in in the power generation, you'll be contributing back into the grid in a way that that benefits ratepayers.

让我们回到AI的用途，以及它是如何改变我们生活的。

Let let's go back to the uses and how AI is changing our lives.

你之前提到了AI的所有用途及其影响。

You you mentioned earlier all of the uses and and and the impact that AI is having.

你认为目前AI最重要、应用最成功的用途是什么？

What do you see as the most important use and where AI is being deployed and implemented best right now?

这很有趣。

Well, it's interesting.

我认为这经历了一个演变过程。

I think there's been an evolution.

我们最初是从像ChatGPT这样的AI聊天机器人开始的。

So I think we started with, you know, AI chatbots like ChatGPT.

从某种意义上说，这就像更高级的网络搜索。

And in a sense, that was kind of like better web search.

它在研究方面非常出色。

It was really great for research.

你可以提出问题，它会为你提供任何问题的答案。

You ask your questions and give you answers to anything.

然后我们看到了具备思维链能力的模型，它们开始能够进行更深入的推理。

Then we saw we saw models at chain of thought and they could start to do, you know, deeper reasoning.

然后我们看到了编程辅助工具。

Then we saw coding assistance.

这是真实存在的，我认为在过去几个月里，如果你去问软件开发者，他们就会觉得有了真正的突破。

This is real and and I think over the past few months, there's been a real breakthrough if you talk to people, software developers.

似乎编程助手的质量有了重大提升和改进。

It really seems like there's been a major shift and just improvement in the quality of the coding assistants.

我认为下一步将是面向知识工作者的工具。

And I think where that's going next is tools for knowledge workers.

因此，那些原本用于输出代码的助手，现在可以输出任何类型的格式。

So the same types of assistants that have been outputting code can now output any type of format.

无论是Excel模型、PowerPoint、网站，还是其他任何东西。

So whether it's like Excel models, PowerPoints, websites, you name it.

知识工作者现在能够像程序员和软件开发者使用AI生成代码一样，生成所有这些不同类型的内容。

Knowledge workers are now gonna be able to generate all these different types of things the same way that coders have been that software developers have been using AI generate code.

我认为2026年我们将看到的一大趋势，就是知识工作者的生产力将迎来新一轮爆发。

I think that's one of the big things we're gonna see in 2026 is, again, this productivity boom for knowledge workers.

所以我认为这就是你在现实中看到的情况之一。

So I think that's like one of the things you're seeing on the ground.

另外，各个行业领域也在发生许多变化。

And then separately, there's there's a bunch of things happening in industry verticals.

不同的行业正受到人工智能的影响。

So different industries being impacted by AI.

在医疗领域，我认为有很大的机会可以减少行政官僚主义，改善繁琐的文书处理流程。

So in health care, I think there's a tremendous opportunity to improve or to to reduce sort of administrative bureaucracy to to improve this processing of paperwork that happens.

同时，还可以利用人工智能在医学和科学研究中帮助发现新的疗法。

Also to use AI and medical and scientific research to help find new cures.

你已经看到用户分享了各种关于诊断的故事。

You're already seeing users tell all sorts of stories about diagnoses.

他们能够将自己的病历输入到ChatGPT或其他聊天引擎或聊天机器人中，并获得惊人的结果。

They've been able to put in their medical records into ChatGPT or your other chat engine or chatbots and get like remarkable results.

他们终于弄清楚了自己到底哪里出了问题，并能带着这些信息去看医生。

They've been able to, you know, finally figure out what was, you know, what what was wrong with them and they've been able to take that to a doctor.

医生们也在使用它。

You have doctors using it too.

所以医疗领域我认为是一个非常有趣的领域。

So medical, I think, is a really interesting area.

但还有很多这样的例子，表明不同行业现在正受到影响。

But there's a whole bunch of these examples of different different industries are now being impacted.

你有吗，是的。

Do you have a Yeah.

我经常思考的一个领域是AI在科学中的应用。

Of the one area I think a lot about is is AI for science.

回到大卫最初关于我们在这些前沿模型上取得进展的观点。

And and back to to to David's initial point about the progress we've seen in these frontier models.

我认为最早的那些模型最初只是基于一般知识。

I think the very early ones sort of started with just general knowledge.

你必须回溯并理解原因，问题是：那些模型构建者在训练模型时可用的数据是什么？

You have to go back and understand like why and the question was, what was the data available for those model builders to start training their models?

对于早期的模型，你只需抓取互联网上的内容，把所有东西都塞进模型里进行训练，这就形成了大型语言模型的第一阶段。

And for the early ones you could just scrape the internet and just kind of cram everything to a model and train it and that's and where you kinda had this this first phase of of large language models.

第二阶段则是编程。

And the second one was coding.

如果你思考如何打造一个优秀的编程模型，你就需要对其进行训练，使用现有的代码进行训练。

And if you think about how do you get a really good coding model, you get you have to trade it, you have to train it on existing code.

这又是一种相对容易获取的数据类型，你因此看到了编程模型的巨大进步和飞跃。

And that's again something that is, you know, relatively easier to to acquire than other types of data and you saw great progress and jumps in in the coding models.

我认为第三个重大转变尚未被充分讨论，而政府本身正在积极推动，那就是科学领域的AI应用。

I think the the third big sort of shift that hasn't really been touched on yet, which the government itself is trying to do a good push on is the AI for science question.

科学发现难以与大型语言模型的传统训练方式结合的原因在于，科学数据极其分散，其形式和格式并不适合直接用于大型语言模型的训练流程。

And why it's so challenging for scientific discovery to like tie in with the way that LMs are are traditionally trained is that the science data is extraordinarily fragmented and it's not done in a way or formatted in a way that can easily be applied to a large language model sort of like training run.

如果你思考科学发现，它分散在众多不同的学科领域中。

And if you think about scientific discovery, it's spread out across so many different disciplines.

你有化学数据、数学数据、材料科学数据，而所有这些数据都采用各种不同的格式。

You have chemistry data, you have math data, you have material science data, and all of that is is all types of different formats.

在我们的政府工作中，我们启动了一项名为‘创世计划’的项目，旨在推动人工智能在科学发现领域的重大突破。

And our effort in administration, we launched something called the Genesis Mission, which are is our attempt to sort of make these big bold leaps in AI for scientific discovery.

能源部下属的国家实验室在过去五六十年间一直进行着卓越的研究，所有这些成果都已准备好，可用于训练这些模型。

And our national labs at the Department of Energy are have been doing incredible research over the last, you know, fifty, sixty years, and all of that has is sitting and is ready to be used to be trained for for for these models.

因此，我希望在未来一年里，我们能在科学发现领域看到更多进展，真正加速我们选择实验、开展实验、回溯错误并重新实验的进程。

So my hope is that over the next year, we're gonna see a lot more work in this in scientific discovery to be able to actually accelerate how quickly we can choose which experiments to run, run those experiments, go back and figure out what we did wrong, and run them again.

这与人们围绕一些AI实验室提出的诸多有趣想法紧密相关——你可以输入假设或理论，这些实验室最终能自主进行实验并推进研究。

And and this ties in with lots of interesting ideas that people have around some of these AI labs where you essentially have you can put in the the the thesis or the hypothesis and ultimately these labs can do lab experiment itself and move forward.

这就是我的愿景：最终，作为一个国家，我们能借助人工智能在未来十年内将研发产出几乎翻倍。

So that's kind of the dream that I have that that ultimately we as a country can can almost double our our r and d output over the next ten years because of AI.

那么，你期望或希望看到什么样的突破呢？

So so what kind of breakthroughs would you expect or would you like to see?

是的。

Yeah.

我认为那些可能产生重大影响的突破首先是围绕核聚变的实验和训练运行，这些过程需要极高的计算量。

I I think they're the ones that I think can make a big impact are, first, the the the the experimentation and training runs around fusion are extraordinarily computation heavy.

如果我们能加快这些聚变模拟的反馈循环，就能缩短聚变技术的实现时间，这将是一个重大突破。

And they themselves, if we can if we can have a faster feedback loop on how we do these these simulations for fusion, we can move the timelines in for fusion, so that could be a big a big step.

材料科学也是一个非常重要的领域，我们需要能够测试各种分子之间的相互作用。

Material science is also a very a very big area where you wanna be able to test all types of of different molecules and interact with each other.

这对我们在太空领域正在推进的所有重大计划都至关重要，无论是月球基地、前往火星，还是将核能引入太空，先进的材料科学都不可或缺。

This is important for all the big things we're trying to do in space, whether it's our lunar base or getting to Mars or bringing nuclear energy to space, having advanced material science is important.

第三个是每个人始终关注的领域：医疗和治疗。

And the third is the one that everyone always cares about is is health care and and therapeutics.

你如何更快地识别出能解决特定健康问题的最佳分子？

How can you more quickly be able to identify the the best molecules to solve a particular particular health challenge?

你如何更快地迭代，直到能够进入临床试验阶段？

And how do you more quickly iterate to a point where you can move to a to a clinical trial?

在日常层面，我认为汽车行业也是这里的主要受益者。

And on an everyday level, I mean, also have the auto sector, I think, as a big beneficiary here.

我认为这是另一个在这一领域投入大量资源的行业。

I think that's one area that seems to be spending a lot on this as well.

你同意这一点吗？

Do you agree with that?

嗯，比如自动驾驶，是的。

Well, I mean, with, like, self driving or yep.

是的。

Yep.

我的意思是，自动驾驶肯定会有巨大发展。

I mean, self driving for sure is gonna be huge.

感觉我们已经到达了一个新的转折点，技术质量已经提升到足以看到机器人出租车了，比如Waymo和特斯拉。

It feels like we've hit some sort of new inflection point there where the quality's gotten to the point where you're starting to see robotaxis now, Waymo's and Tesla.

那AI助手呢？

What what about an AI assistant?

我的意思是，这会成为一种普遍存在的东西吗？

I mean, is that gonna be something that is sort of commonplace?

前几天有人跟我说，在中国，你们的做法完全不同，因为你们用AI做研究，就像你所说的，但我们在用它当我的AI助手，帮我还账、打扫房子、给我妻子买生日礼物，什么都替我做了。

Someone said to me the other day that, oh, in China, we're doing things so much differently because you're using AI for research, as as you said, but we're using it as I have my AI assistant, and, you know, they're paying my bills and cleaning my house and buying my wife a birthday present and and doing everything for me.

我觉得是的。

I I think so.

我认为今年就会发生。

I think that'll happen probably this year.

最近刚刚发布、让所有人都疯狂追捧的产品是Cloud Code的最新版本，它由Anthropic的Opus 4.5模型驱动，这在编程领域似乎是一次真正的突破。

So the the the product that just came out recently that everyone's kind of going crazy over is the latest iteration of Cloud Code, which is powered by Anthropix Opus 4.5 model, which seems to be a real breakthrough in in in coding.

所以，同样地，软件开发者们对它印象深刻。

And so, again, this is you know, the soft developers are really impressed with it.

但在Claude Co内部，他们推出了一项名为‘协作’的新标签，在这里，即使你不是程序员，或者你希望生成的不是代码，也可以用它来创建各种其他类型的内容。

But in inside of Claude Co, they had they introduced a new tab called co work where, again, you can as a non coder or as someone who is looking for to create output other than code, you can now use it to to basically create all sorts of other kinds of outputs.

就像我提到的，你可以用它做电子表格或PowerPoint演示文稿之类的东西。

Like I mentioned, you can do spreadsheets or PowerPoints, things like that.

而且你可以让它。

And you can have it.

你可以让它访问你的文件驱动器，查看你之前做过的工作。

You can point it to your file drive and it can look at the work you've already done.

所以，如果你喜欢某种特定的PPT格式，只需指向你之前的工作内容，然后说：我想做一个新的演示文稿，但用这种风格，主题换成这个。

So if there's a particular type of format for a PowerPoint you like, you just point it to the work you've already done and say, I wanna do, you know, a new presentation but using this style but on this topic.

它会真正模仿你的风格、格式以及你之前的工作成果。

And it'll actually emulate, you know, your style and and the work, your format, the work you've already done.

而且人们对这一点非常印象深刻。

And and people are very impressed with this.

你还可以让它连接到你的邮箱，让它分析你的邮件并从中提取信息。

And you can also point it at your email and have it analyze your email, pull things out of it.

所以目前它非常以任务为导向。

So it right now, it's very task based.

你需要为每个任务主动向它发出指令。

You you, the user, have to prompt it for each task.

但你可以看到，一个个人数字助手的雏形已经出现：你可以将它连接到你的文件夹、邮箱和所有数据源，让它开始为你执行任务。

But you can see there the beginning of a personal digital assistant where you connect it to your file drive, to your email, to all of your data sources, and it can start to do tasks for you.

而且它还能理解你偏好的工作格式和风格。

And again, it understands the format and the style that you like to produce work in.

所以在我看来，我们只需要在这样的工具之上再加一层抽象。

So it feels to me like we just need one more layer of abstraction on top of a tool like that.

那你就会拥有自己的个人数字助手。

And you'll have your own personal digital assistant.

而且，会有一个语音界面。

And you know, there'll be like a voice interface.

你看过电影《她》吗？

You ever seen the movie Her?

你知道的，由华金·菲尼克斯主演，斯嘉丽·约翰逊只提供声音。

You know, with Joaquin Phoenix and Scarlett Johansson is just the voice.

但你知道，他是通过耳机告诉她该做什么。

But you know, he's telling her what what to do through an earpiece.

我们的距离已经非常接近这样的场景了。

I mean, we're very close to something like that.

我的意思不是说，AI会变得有意识之类的。

I mean, I'm not saying that, you know, the AI is gonna become sentient or whatever.

但不，不是这样的。

But but no.

我认为到2026年，你就能看到这类工具——最初是编程助手，但现在正逐渐演变为个人数字助手。

We're like I think in 2026, you could see that that these types of of tools, again, started as coding assistants, but now they become personal digital assistants.

今年就完全有可能实现。

That could definitely happen this year.

迈克尔，人们不理解AI的方面有哪些？

Michael, what what don't people understand about AI?

你认为我们现在正在发生的科学与AI创新中，最需要我们理解的是什么？

What what do you think is most important for us to understand about the innovation underway right now with science and and AI?

我认为有些人很容易低估AI在未来对如此多行业和领域产生的长期影响。

I I think some people I I think it's easy to underestimate the the long term impact this is gonna have across so many industries and and domains.

我觉得人们很容易把AI简单地看作一个高级聊天机器人，因为大多数人每天接触和体验的正是这种形式。

I think very much, you know, it's easy to to to quickly think about AI as a as just a sophisticated chatbot because that's what most people interact with every day and and that's what they they they touch and feel.

但在我看来，长远的影响在于——我不想老是扯到科学上，但确实正在发生一场根本性的变革，那就是我们测试、评估和推进科学发现的速度和节奏正在发生根本性变化。

But I think that to me, I think the long term impacts and not to keep harping on the science, I think there is a there's a a real fundamental shift happening in the velocity and pace that we can test and and evaluate and execute scientific discovery and endeavors.

我认为这将对我国在未来几年的创新方式产生巨大影响，从广义上讲。

And I think I think that's gonna have huge repercussions for the way that we as a country innovate, broadly speaking, in the years ahead.

这就是为什么我们在关注中国在做什么。

Which is why we're watching what China is doing.

让我们谈谈中国，以及它相对于美国的位置。

Let's talk a bit about China and where it is relative to The United States.

我们赢了吗？

Are we winning?

这关乎芯片吗？

Is it about chips?

这场竞赛究竟真正关乎什么？

What's the race specifically really about?

我认为总体而言，我们领先于中国。

Well, I I I think that in general, we're ahead of China.

这个技术栈有不同的层面。

There's different layers of the stack.

所以你有模型，然后有芯片，再往下是芯片制造设备。

So you've got the the the models, then you've got the chips, and, you know, then you've got the chip making equipment.

你知道的，你沿着这个技术栈往下看。

You know, so you you go down the stack.

我认为，你往技术栈的底层走，美国的优势就越大。

I would say that the deeper in the in the stack that you go, the greater the American advantage.

我觉得在模型方面，大多数人会说我们的模型比中国模型领先大约六个月左右，上下浮动。

I think on models, most people would say that we're our models are maybe six months ahead or so, plus or minus, the Chinese models.

你看看芯片，可能领先两年。

You look at chips maybe two years ahead.

到了半导体制造设备，差距可能达到五年左右。

You go to the semiconductor manufacturing equipment, it could be, like, five years.

所以美国确实拥有显著的优势。

So The US does have significant advantages.

只有少数几个领域，我认为中国可能占优势。

There there's only maybe a couple areas where I think China has has an advantage.

一个是能源生产。

One is on energy production.

如果你看一下他们的电网，过去十年里他们的电网规模大约翻了一番。

If you look at the their grid, their grid has roughly doubled in the last ten years.

而我们的电网仅增长了大约百分之二到三。

Whereas ours has only grown by about two to 3%.

在人工智能出现之前，美国的能源生产一直是一个相对沉寂的行业。

Energy production in The US has been a relatively sleepy industry before AI came along.

这很大程度上是因为监管限制以及前一届政府对能源生产的抵触态度。

And a lot of that had to do with regulations and the antipathy of the previous administration towards energy production.

显然，特朗普总统对此持非常不同的看法。

Obviously, president Trump had a very different view on this.

我认为他在这一问题上很有远见。

I think he was prescient on this issue.

十年前他就说过：‘钻吧，宝贝，继续钻。’

You go back ten years and he was talking about we gotta drill, baby, drill.

我认为他明白，能源增长是经济发展的前提，也绝对是AI基础设施发展的前提。

And and I think he understood that energy growth was the precondition for economic growth and is definitely the precondition for this AI infrastructure growth.

因此，这一领域我们再次必须大幅扩大能源生产。

So this is an area where, again, again, we have to basically expand our energy production.

我认为这是我们亟需追赶的领域。

And I and and and and so I think that is an area where we need to catch up.

另一个领域，我想说，我不知道是否能称之为优势，但你可以说中国在所谓的AI乐观主义方面占据优势。

The other area where I would say, you know, I don't know if I would call this an advantage exactly, but if you but you could argue that China China has the edge in what is what's being called AI optimism.

斯坦福大学对各国进行了一项民意调查。

So there was a a polling done by Stanford across countries.

他们询问这些国家的公民：你认为AI的好处是更多还是危害更大？

And they asked the citizens of all these different countries, do you feel that the benefits of AI will be more beneficial or more harmful?

如果你认为AI总体上利大于弊，他们就称之为AI乐观主义。

And if if you thought that it that overall be more beneficial than harmful, they call that AI optimism.

在中国，AI乐观主义的比例达到了83%。

Well, in China, AI optimism was 83%.

所以有83%的人认为AI带来的好处多于害处。

So 83% of the population feels that it's being more beneficial than harmful.

在美国，这个数字只有39%。

That number in The United States is only 39%.

因此，不知为何，中国人对AI的乐观程度高于美国人。

So for some reason, people in China are more optimistic about AI than in The United States.

通常你会看到，亚洲国家的AI乐观指数普遍很高，而西方国家则较低。

And you generally you generally see this that Asian countries are very high on AI optimism, the Western countries are lower.

我认为这是一个有趣且开放的问题，即为什么会这样。

And I think it's a interesting or open question about why this is.

我认为有几个可能的解释。

I think there's a few possible explanations for it.

我认为，首先，媒体往往更关注AI的悲观和负面故事。

I I think that, first of all, the the media tends to focus on the doom and gloom stories with with AI.

恐惧。

The fear.

这些恐惧。

The fears.

我们可以谈谈其中一些恐惧，以及你认为它们是否真实。

And we can talk about some of those fears and and and how, you know, whether we think they're they're real.

但我认为媒体在这方面起了很大作用。

But I think the media has a lot to do with it.

我认为，好莱坞几十年来对人工智能的描绘——无论是《终结者》还是《2001太空漫游》——都呈现了一种反乌托邦的未来图景。

I think that the the way that Hollywood has portrayed AI over the decades, you know, with whether it's the Terminator or 2001 has, you know, has portrayed this dystopian view of the future.

我认为这影响了博尔的思维。

And I think that places Buell's thinking.

坦率地说，我认为部分责任在于我们的科技领袖，他们并没有很好地阐述人工智能的好处。

And then frankly, I would say that part of the the fault lies with our tech leaders who haven't necessarily done a great job describing the benefits of AI.

事实上，当他们谈论人工智能将取代50%的知识型工作者时，这听起来并不像一个美好的愿景。

In fact, when they're talking about, you know, AI eliminating 50% of knowledge workers, that doesn't sound like a, you know, very utopian scenario.

对大多数人来说，这听起来像是一个反乌托邦的场景。

That sounds dystopian to most people.

因此，我认为我们的某些科技领袖无意中助长了这种对AI的悲观情绪。

And so I do think that unintentionally, some of our tech leaders have played into this AI pessimism.

我认为这可能对美国构成劣势的原因是，它进一步加剧了我们正在经历的监管狂潮。

And the reason why I think this could be a disadvantage for The United States is because, again, it's feeding into this regulatory frenzy we're seeing.

目前，各州有1200项相关法案。

Again, 1,200 bills at the state level.

而目前，我认为我们在AI竞赛中处于领先地位。

And right now, I think, you know, we are winning this AI race.

我们在芯片、模型等所有关键领域都遥遥领先。

We're ahead in all the key dimensions, chips, models, and so on.

但我们可能会自毁前程。

But we could shoot ourselves in the foot.

你知道吗？

You know?

如果我们最终对这项技术监管过度，反而可能让我们输掉这场AI竞赛。

If we end up overregulating this thing to to death, we could actually cost ourselves this AI race.

所以我确实担心这个关于AI乐观主义的问题。

So I do worry about this question of AI optimism.

没错。

Right.

这是个很好的观点。

It's a great point.

如果美国在这场竞争中不是第一，会发生什么，迈克尔？

And what would happen if The US is not number one in this, Michael?

是的。

Yeah.

我认为我们必须如此，这正是我提出这个计划的原因。

I I I think we we need to be and that's why I put put the plan out.

当我思考中国问题以及更广泛的议题——我们如何赢得AI竞赛时，我认为……

I think, you know, when when I think about the the China question and about the the sort of larger question, how do we win the AI race?

我总是喜欢思考这个问题：采用率。

What I always what I always like to think about is this question of adoption.

我认为有时人们过于强调排行榜了。

And I think sometimes there's this overemphasis on the leaderboard.

就好像哪个前沿模型在某种指标上排第一。

It's like which frontier model is number one on some sort sort of metric.

但事实上，我们双方势均力敌，正如大卫所说，我们在前沿模型上可能只领先六个月到一年。

And the reality is we're neck and neck and as David said, we'll probably had, you know, six to twelve months in our frontier models.

但纵观历史，我们发现，你并不一定需要拥有世界上最好的模型或最顶尖的技术，才能让其在全球普及。

But I think what we have seen over over time and over history is that you don't necessarily need to have the very best model or very best piece of technology in the world for it to proliferate globally.

我们许多曾参与第一任特朗普政府的人，都亲身经历了那个时代的电信战争，见证了华为在全球的作为。

And a lot of us who were part of the first Trump administration saw this very firsthand with the telecom wars of that era of what Huawei was able to do globally.

当华为刚开始推动全球出口时，他们的技术绝对不是世界上最好的。

And at the time when when Huawei first started their their sort of global export push, they certainly were not the very best technology in the world.

他们的技术在当时确实不如爱立信和诺基亚，但足够好，而且得到了充分补贴，因此成为了全球许多地区的默认电信系统。

They were current they were certainly, you know, you know, subpar compared to to Ericsson and Nokia, yet they were good enough and they were subsidized enough such that they became sort of the the default telecom system for a lot of the world.

我们从中学到了很多教训，在人工智能领域，我们对此非常重视。

And we've learned a lot of lessons from that, and we take that very seriously when it comes to AI.